Intelligent license plate recognition method

Abstract

The invention relates to the technical field of intelligent license plate recognition, in particular to an intelligent license plate recognition method. The method comprises the following steps: firstly, acquiring multiple frames of images from a video stream, and detecting and caching effective vehicle images to form a sequence; detecting a license plate area in the sequence by using a YOLOv5 network, and intercepting the area after confidence filtering and duplicate removal to form a candidate set; performing multi-dimensional quality evaluation and fusion calculation on the candidate images, and preferably selecting a high-quality license plate image according to the comprehensive score; carrying out character segmentation on the optimized image, and carrying out quality evaluation and classification on each character; qualified characters are directly identified, fuzzy characters are firstly subjected to image enhancement and then identified, and defective characters are inferred and identified in combination with coding rules; and assembling the identification results into a complete license plate number and calculating the overall confidence coefficient, if the overall confidence coefficient does not reach a threshold value, triggering a feedback loop, enhancing related images, and re-evaluating and identifying the related images. The method effectively improves the accuracy and efficiency of license plate recognition under motion blur.

Description

Technical Field

[0001] This invention relates to the field of intelligent license plate recognition technology, and more specifically, to an intelligent license plate recognition method. Background Technology

[0002] License plate recognition technology is a crucial component of intelligent transportation systems for vehicle identification. However, in real-world checkpoint and road monitoring scenarios, the high-speed movement of vehicles often results in severe motion blur in the captured license plate images. This leads to a significant loss of edge and texture details in the characters, becoming one of the primary challenges affecting recognition accuracy.

[0003] To solve the above problems, existing technologies mainly adopt the following two solutions: One approach is to use an end-to-end deep recognition network. This approach directly inputs the entire blurred license plate image into a unified convolutional neural network model. Through end-to-end training, the model can directly output the recognition result from the input blurred image. Second, deep learning-based image restoration technology is used. First, a deblurring or super-resolution model is used to perform global enhancement processing on the entire license plate image, and then the enhanced image is sent into the standard license plate recognition model.

[0004] However, the two schemes mentioned above have certain limitations: for end-to-end recognition schemes, when the input image has severe motion blur, the model lacks the ability to perceive the degree of blur and the targeted processing mechanism, so its ability to extract effective features from highly distorted images is limited, resulting in a significant decrease in recognition accuracy. The global enhancement scheme's computational mode of uniformly processing the entire license plate image not only leads to high computational overhead and processing latency, but more importantly, it cannot distinguish the degree of blur in different regions of the image, resulting in unnecessary enhancement operations on originally clear character regions.

[0005] Therefore, there is an urgent need for a license plate recognition method that can intelligently perceive image quality and perform differentiated processing accordingly, thereby improving processing efficiency while ensuring recognition accuracy. Summary of the Invention

[0006] This invention provides an intelligent license plate recognition method, which involves acquiring multiple frames of images from a video stream and filtering out valid license plate regions; performing multi-dimensional quality assessment on the images to select high-quality license plates; segmenting and independently assessing and classifying the license plate characters; performing differentiated recognition based on the character quality classification results; and enhancing and reprocessing low-quality images through a confidence-driven feedback loop, thereby solving the problems mentioned in the background art, namely: In complex motion scenarios, license plate images suffer from motion blur, leading to a decrease in recognition accuracy and a difficulty in balancing processing efficiency.

[0007] To achieve the above objectives, the intelligent license plate recognition method includes the following steps: S1. Acquire multiple frames of images from the video stream, cache valid vehicle images after vehicle presence detection, and generate a multi-frame image sequence; S2, Use The network performs license plate region detection on a multi-frame image sequence. After confidence filtering and non-maximum suppression, the license plate region is extracted from each frame to form a candidate set of license plate region images. S3. Perform multi-dimensional quality assessment on each image in the candidate set of license plate area images, standardize the scores of each dimension, and calculate the comprehensive quality score by weighted fusion. Based on the score, select one or more high-quality license plate images from the candidate set. S4. Perform grayscale and adaptive binarization processing on the selected license plate images, locate character boundaries through vertical projection analysis, complete character segmentation to obtain individual character images, and conduct quality assessment on them, classifying characters into qualified characters, blurred characters, and missing characters. S5. Standard recognition is used for qualified characters, image-enhanced recognition is used for blurred characters, and contextual reasoning combined with encoding rules is used for missing characters. The recognition results of each character and their confidence scores are assembled into a complete license plate number in sequence, the overall confidence score is calculated, and a success threshold is set. If the overall confidence level reaches the preset success threshold, the complete license plate recognition result is output. If the success threshold is not reached, motion blur recovery enhancement based on convolutional neural network is performed on the original license plate image corresponding to the low confidence character, and steps S3 to S5 are repeated.

[0008] The above technical solution employs a multi-frame acquisition and caching initial step to build temporal data redundancy, providing flexibility for subsequent stages. Relying solely on single-frame information would render the system incapable of responding to sudden drops in image quality. Introducing multi-dimensional quality assessment and comprehensive score calculation quantifies image usability from multiple perspectives, avoiding misselection due to a single quality defect. Relying solely on a single sharpness index might overlook issues caused by lighting, occlusion, or deformation. Character-level segmentation and classification based on sharpness and integrity are based on the core design principle that different characters on a license plate may exhibit varying degrees of degradation, necessitating differentiated processing strategies. Traditional methods of uniformly enhancing or recognizing the entire license plate image cannot effectively enhance blurred areas, rapidly process sharp areas, or logically supplement missing areas. The combination of differentiated recognition channels and a confidence-based feedback loop forms a closed-loop system of evaluation-processing-verification-optimization, allowing resources to be concentrated on the most critical stages. Without such mechanisms, the system would be unable to self-correct and improve when recognition confidence is insufficient.

[0009] Based on this, the The network enhances its ability to detect license plate targets at multiple scales by introducing a feature pyramid network module into its structure, thereby optimizing the localization accuracy for small-sized and tilted license plates.

[0010] In another technical solution, when inputting the image into the... Before processing the image, a fill scaling process is performed to maintain the aspect ratio, adjusting the image to the network input size and avoiding geometric distortion.

[0011] This technical solution specifically introduces a feature pyramid network module into the license plate region detection step. Its design aims to enable the network to simultaneously utilize the detailed information of shallow features and the semantic information of deep features, effectively addressing the differences in license plate scale caused by variations in shooting distance and the tilting issues caused by changes in viewing angle. Without this multi-scale feature fusion mechanism, the network's ability to detect small or tilted license plates would significantly decrease, easily leading to missed detections or inaccurate bounding boxes. Simultaneously, a fill scaling operation that maintains the aspect ratio is used during input image preprocessing. This is to preserve the original proportions and geometric features of the license plate region when adapting images of arbitrary sizes to the network's fixed input size, avoiding shape distortion caused by direct stretching or compression. Simple deformation scaling would distort the character shapes in the license plate, thus interfering with the detection network's judgment of the license plate boundary and angle. These two improvements work synergistically from two levels: optimizing the internal network structure and preserving external input fidelity, jointly enhancing the adaptability and accuracy of the license plate region detection process.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by introducing a character-level quality assessment and classification mechanism, achieves for the first time in the license plate recognition process independent diagnosis and differentiated processing of individual character states. This design enables the system to accurately distinguish between clear, blurred, and missing characters, and allocate different recognition resources accordingly: clear characters are directly recognized to ensure speed, blurred characters are enhanced before recognition to improve accuracy, and missing characters are inferred using contextual rules to compensate for insufficient visual information. This refined, heterogeneous processing mode fundamentally avoids the unnecessary processing of clear areas by traditional global enhancement methods. While significantly reducing average computational overhead, it ensures that processing resources are concentrated on the blurred or missing areas that most need improvement, thereby achieving synergistic optimization of recognition efficiency and accuracy.

[0013] 2. This invention constructs a closed-loop feedback optimization system based on confidence assessment. This design uses the overall confidence level of the recognition results as a quantitative indicator to measure the success or failure of a single recognition attempt, driving a dynamic feedback loop: when the confidence level is insufficient, the system does not fail directly, but automatically locates the original image corresponding to the low-confidence character, triggers a targeted image enhancement process, and re-executes subsequent quality assessment and recognition steps. This closed-loop mechanism endows the system with the ability to "self-optimize," enabling it to iteratively enhance and re-recognize difficult samples that fail on the first attempt, thereby continuously improving the final recognition success rate and result reliability for complex scenes such as severe motion blur. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the overall structure of the intelligent license plate recognition method of the present invention. Figure 2 For the present invention Network structure diagram; Figure 3 This is a flowchart of step S5 of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Currently, to address the problem of decreased recognition accuracy and difficulty in balancing processing efficiency caused by motion blur in license plate images during complex motion scenarios, this invention provides an intelligent license plate recognition method. (See [link to relevant documentation]). Figure 1 As shown, it includes the following steps: S1. Acquire multiple frames of images from the video stream, cache valid vehicle images after vehicle presence detection, and generate a multi-frame image sequence; S2, Use The network performs license plate region detection on a multi-frame image sequence. After confidence filtering and non-maximum suppression, the license plate region is extracted from each frame to form a candidate set of license plate region images. S3. Perform multi-dimensional quality assessment on each image in the candidate set of license plate area images, standardize the scores of each dimension, and calculate the comprehensive quality score by weighted fusion. Based on the score, select one or more high-quality license plate images from the candidate set. S4. Perform grayscale and adaptive binarization processing on the selected license plate images, locate character boundaries through vertical projection analysis, complete character segmentation to obtain individual character images, and conduct quality assessment on them, classifying characters into qualified characters, blurred characters, and missing characters. S5. Standard recognition is used for qualified characters, image-enhanced recognition is used for blurred characters, and contextual reasoning combined with encoding rules is used for missing characters. The recognition results of each character and their confidence scores are assembled into a complete license plate number in sequence, the overall confidence score is calculated, and a success threshold is set. If the overall confidence level reaches the preset success threshold, the complete license plate recognition result is output. If the success threshold is not reached, motion blur recovery enhancement based on convolutional neural network is performed on the original license plate image corresponding to the low confidence character, and steps S3 to S5 are repeated.

[0017] Through the collaborative working mechanism of multi-frame acquisition, quality assessment, character-level processing, and feedback verification, accurate and stable recognition of license plate characters is achieved in complex motion scenarios, effectively improving the reliability and robustness of the license plate recognition system in practical applications.

[0018] Step S1 of this invention involves multi-frame image acquisition and buffering. This step aims to construct a temporally coherent multi-frame image sequence from a continuous surveillance video stream, containing the complete visible portion of the target vehicle, thus providing a high-quality data foundation for subsequent processing. The input is the original video stream, and the output is a structured data packet containing N valid vehicle images, i.e., a multi-frame image sequence. This is specifically achieved through the following sub-steps.

[0019] Ⅰ: Establish a complete data receiving channel and initialize the processing environment. Its core task is to convert the raw video data from the outside into a standardized data stream that the system can process, and to create an ordered buffer space for subsequent processing.

[0020] ① First, establish a physical connection with the video source. Let the video source address be... A connection channel is established through a network protocol handshake. : In the formula, This represents a channel object indicating a successful connection to the video source. It is a string variable representing the network address of the video source; This is a time parameter that represents the maximum waiting time for the system to attempt to establish a connection. If the connection is not established successfully after this time, it is considered a connection failure.

[0021] After a successful connection, obtain the basic parameter set of the video stream: In the formula, It is a set of parameters that contains key information describing the characteristics of the video stream; For frame rate; These represent the original width and height of the video stream, i.e., the video resolution; Indicates the color space format of the video; This indicates the video bitrate.

[0022] ② Initialize the corresponding decoder instance based on the obtained video parameters. : In the formula, This represents a decoder instance object that has completed initialization; This is an initialization function, which is based on the video. and Configure and start the corresponding decoder.

[0023] Establish a decoding frame buffer Its size matches the video resolution: Simultaneously configure a color space converter to convert the original video color space. Unified conversion to standard space: In the formula, Represents a color space converter object; Used according to the source color space and target color space Create a mapping or function for conversion between the two.

[0024] ③ Create a first-in-first-out (FIFO) cache queue using the following formula. : In the formula, It is the array that stores the queue. Each element can store a single image frame. It is an integer constant representing the maximum capacity of the queue; It is the state vector of the queue.

[0025] Each storage unit Allocate fixed memory: In the formula, This indicates that each buffer is calculated. The amount of memory required; It is the total number of pixels in the image; Indicates the number of RGB color channels (red, green, blue); This indicates the number of bytes occupied by a single pixel value in each channel.

[0026] ④ Constructing the video timeline With processing time axis The mapping relationship. Let the timestamp of the first frame of the video be... The system receiving time is Then establish a linear mapping: Simultaneously create a timestamp array Used to record the capture time of each frame.

[0027] After completing the basic video streaming environment setup, set the minimum acceptable quality score. This parameter, based on a comprehensive evaluation of image sharpness, contrast, and noise level, is used to remove invalid frames caused by transmission errors or severe motion blur. Next, an inter-frame similarity threshold is defined. By calculating the structural similarity index of consecutive frames, a large number of duplicate images are avoided when the vehicle is stationary or moving slowly. Finally, the maximum number of consecutive dropped frames is configured. When the system drops more than this threshold of consecutive frames due to substandard quality, a video stream health check mechanism will be triggered. These three parameters together constitute the baseline conditions for quality control, ensuring that subsequently acquired image sequences meet both recognition requirements and maintain temporal validity.

[0028] II: Based on the already initialized video processing environment, perform regular frame acquisition and standardization processing. Its core task is to extract image frames from the video stream at preset time intervals and convert these frames into a unified format required by subsequent processing.

[0029] First, establish a strict time control mechanism with fixed time intervals. Trigger frame acquisition. Assume the system clock is... The last sampling time was The sampling conditions are: Frame acquisition is performed immediately when this condition is met, and updates are made. .

[0030] Further, frame decoding and format conversion are performed. The acquired compressed data packet is sent to the initialized decoder for decoding to restore the original image frame. Then, the configured color space converter is used to uniformly convert the image from the original color space to the standard RGB format to ensure consistent color representation.

[0031] After color conversion, image size normalization is performed. Regardless of the original video resolution, the image is scaled to a system-preset standard size. This process uses a bilinear interpolation algorithm to maintain sufficient image quality while ensuring processing speed. Specifically, for RGB input images... Standardize the size by scaling it to the system's preset standard size. : In the formula, Indicates the output image; This is the image scaling function.

[0032] The steps of using the bilinear interpolation algorithm are as follows: In the formula, This indicates the coordinates on the output image. The color value of the target pixel; For the original input image The corresponding coordinates for sampling; These are the interpolation weights.

[0033] This operation maps pixel values ​​to the [0,1] interval, which is beneficial for subsequent neural network processing.

[0034] Each processed image is bound to corresponding metadata, including acquisition timestamp, frame sequence number, original resolution information, and normalization parameters, forming a complete normalized image dataset.

[0035] III: Receive a standardized image dataset and determine whether the frame image contains a complete vehicle target that can be used for license plate recognition, and decide whether to include it in the cache sequence for subsequent processing.

[0036] First, standardized images are extracted from the standardized image dataset and fed into a lightweight vehicle presence detection module. This module is an optimized binary classification convolutional neural network with a streamlined network structure to ensure inference is completed in milliseconds. The network is trained to focus on the center and neighborhood regions of the image, analyzing whether a complete, unobstructed vehicle target exists within these regions.

[0037] The final layer of the network outputs an unnormalized real value. To convert this value into a confidence score representing a probability, the network uses... Activation function The processing is performed, and its mathematical expression is: In the formula, Indicates the confidence score; This represents the raw score output by the final layer of the vehicle detection network. It is a natural constant; To perform exponential operations after negating the original score; for function.

[0038] This function will accept any real value. Mapped to the interval (0, 1), output value This is the confidence score; the closer it is to 1, the higher the probability that a vehicle exists.

[0039] At the same time, based on a preset judgment threshold Generate a binary decision value : Next, the system according to and Perform a joint determination. If and only if... That is, the network believes that a vehicle exists and In other words, only when the certainty of the judgment is high enough will the current frame be officially marked as a valid vehicle image. This image and all its associated metadata will then be allowed to proceed to the next step for caching. If any of the above conditions are not met, the current frame is considered an invalid frame, and the system immediately discards it, thus filtering out background images without vehicles or frames with incomplete vehicle targets.

[0040] IV: This step is initiated only when a frame data marked as a valid vehicle image is received. The core task of this step is to maintain a first-in, first-out (FIFO) buffer queue to ensure that the queue always contains the latest consecutively captured N high-quality vehicle images.

[0041] First, a fixed-capacity first-in-first-out (FIFO) buffer queue is established. This queue is implemented as a circular buffer in memory and is specifically used to store vehicle images that have passed the validity determination and their metadata.

[0042] Queue management uses a circular buffer structure, through state vectors. It accurately describes the real-time status of the queue. Among other things... Record the current number of valid frames. Pointing to the oldest frame position, Points to the next write position. When a new valid frame arrives, the system executes the following decision process:

[0043] If

[0044] mod

[0045] Else (Queue is full):

[0046] (Covering the oldest frame)

[0047] mod

[0048] mod

[0049] First, check the queue status. If If the queue is not full, the system will store the current frame. Point to the location, then press mod Update the tail pointer, and through Increment the count. If If the queue is full, an overwrite operation is performed: the new frame is stored. The pointed-to location overwrites the oldest frame, then press... mod and mod Update the head and tail pointers synchronously.

[0050] After completing the above steps, the system outputs the updated cache queue and its corresponding status information. This continuously maintained image sequence provides a stable and reliable data source for the generation of subsequent multi-frame image sequences, ensuring that subsequent processing steps can be analyzed and identified based on a set of temporally continuous and quality-filtered image frames.

[0051] V: The monitoring system triggers a signal. When a recognition request is received or the vehicle's location reaches the optimal recognition area, a snapshot operation is immediately performed on the current cache queue.

[0052] The system employs a dual-trigger mechanism to ensure accurate output timing: firstly, it responds to recognition requests proactively sent by subsequent processing modules; secondly, it automatically determines whether the vehicle has entered the optimal shooting area through built-in trajectory analysis logic. When either condition is met, the system immediately pauses queue updates to maintain the stability of the current cache state.

[0053] After locking the queue, the system executes the data extraction process according to the image time sequence. Considering the circular storage structure of the queue, the reading process needs to perform a circular traversal based on the head pointer position in the state vector and the number of valid frames to ensure that the output frame sequence maintains a strict chronological order. Starting from the position indicated by the head pointer, the system reads the consecutively stored valid frames sequentially. When it reaches the end of the queue, it automatically loops back to the beginning of the queue to continue reading until all valid frames have been acquired.

[0054] After completing the frame data reading, the system integrates and encapsulates the ordered image frames with their corresponding metadata information. Each image frame retains complete metadata such as the acquisition timestamp, sequence number, and resolution parameters. This information, together with the image data, constitutes a complete data unit. Finally, the system packages all data units into a structured multi-frame image sequence data packet and transmits it to the subsequent license plate area localization processing module.

[0055] Through the standardized output process in this step, the dynamically acquired vehicle images are transformed into structured batch data, providing a high-quality, time-series-guaranteed input source for subsequent processing stages, thus ensuring the stability and reliability of the entire recognition system's data flow.

[0056] Step S2 of this invention is license plate region localization and preliminary screening. Specifically, the core task of this step is to comprehensively analyze the multi-frame image sequence data packets from S1, accurately locate the possible license plate regions in each frame through an optimized target detection network, and screen the localization results based on detection quality indicators, finally outputting a set of high-quality license plate region image candidates, providing an accurate input data source for subsequent processing.

[0057] I: The input continuous image sequence is deconstructed into independent processing units, and precise license plate region detection is performed on each frame, constructing a complete mapping from sequence data to single-frame detection results. First, the multi-frame image sequence is iteratively deconstructed, and each frame is extracted sequentially in chronological order.

[0058] For each frame of image, perform the following standardized detection process: ①The original input image Adjust to the input size required by the detection network. This process uses a fill scaling method that maintains the aspect ratio: In the formula, The processed image; A fill scaling function to maintain aspect ratio.

[0059] By employing a fill scaling method that maintains the aspect ratio, image distortion is avoided, ensuring that the geometric features of the license plate area are not destroyed.

[0060] ② Input the processed image into the improved... The detection network enhances its ability to detect license plates at different scales through a feature pyramid. The network output contains three key components: In the formula, Indicates the network's response to the first... The complete output of the frame; The bounding box matrix; This is the confidence vector; This is a category vector.

[0061] This network, through its enhanced feature pyramid structure, can effectively detect license plate targets of different scales and perspectives, with particularly optimized detection capabilities for small-sized and tilted license plates.

[0062] like Figure 2 As shown, The network structure mainly consists of three parts: the backbone network, the neck network, and the head network, each with a unique function.

[0063] The Backbone module is responsible for extracting features from the input image. It uses CSPDarknet53 as the backbone network. This network adopts the approach of introducing a Cross-Stage Partial Network structure to effectively reduce computational redundancy and enhance feature extraction efficiency. CSPDarknet53 relies on multiple convolutional layers and residual connections to extract a large amount of feature information from the input image and achieve multi-scale presentation. It is mainly used to obtain low-level features in the image.

[0064] The Neck module performs multi-stage fusion and enhancement of features extracted from the Backbone. This module is built upon two structures: a feature pyramid network and a path aggregation network. The feature pyramid network, through top-down path fusion of feature maps at various scales, improves the model's detection capabilities for targets at multiple scales. The path aggregation network, by strengthening the path aggregation module, enhances cross-level feature transmission and information interaction, allowing the network to better preserve and integrate key information in complex backgrounds and scenarios where targets are occluded. This design allows... Achieving a balance between accuracy and efficiency when detecting targets of different sizes.

[0065] The Head module is responsible for generating the final detection results, namely the target category and bounding box coordinates. This module processes the feature map output by the Neck to generate target category predictions and bounding box regression predictions at each location. The Head uses feature maps at three different scales for multi-scale prediction, with each scale responsible for detecting targets of a specific size, maintaining high accuracy across different scales, especially excelling in small target detection. The Head's output includes the class probability of each predicted box, the bounding box position with center coordinates, width and height, and offset relative to the anchor box, and a confidence score.

[0066] ③ Map the normalized bounding box coordinates output by the network back to the original image coordinate system: In the formula, The normalized coordinates are mapped back to the actual pixel coordinates in the original image coordinate system, which are the original image coordinates. For the first Normalized coordinates of each detection box; It is the inverse coordinate transformation function; The scaling parameter records the scaling ratio from the original image to the input network; For fill parameters, record the fill pixel information added when resizing.

[0067] This transformation is based on scaling and padding parameters during preprocessing, ensuring that the detection box accurately corresponds to its actual position in the original image, providing precise coordinates for subsequent image cropping.

[0068] After obtaining the detection results output by the network, the coordinate information based on the standardized input image needs to be accurately converted back to the actual pixel coordinate system of the original image. The system performs inverse calculations based on the image scaling ratio and added padding region information recorded during the preprocessing stage, as well as the coordinates of each detection box. At this point, the system remaps the position of each detection box to the pixel space of the original input image, ensuring that subsequent steps can extract the target region from the original image based on the correct spatial location, thereby guaranteeing that the final license plate candidate image has an accurate geometric correspondence.

[0069] ④ Format the detection results of each frame into a uniform structure: In the formula, For the first The set of frame detection results; The data structure for a single detection result.

[0070] Each detection result includes bounding box coordinates, confidence score, and class label. This structured encapsulation is used for subsequent steps to uniformly filter and process the detection results.

[0071] II: To filter out low-quality detections, a global confidence threshold is set. This threshold is determined based on a large amount of validation data to distinguish reliable detections from noise or false detections. For each frame of detection results in the original detection result set, each detection candidate is traversed. This filtering process can be formally represented as constructing a new set: In the formula, Indicates the first The set of detection results retained after the frame image is filtered by a confidence threshold; It is the original, single detection result data structure, from the first... The first frame One detection candidate; It is a candidate for detection The corresponding confidence score; It is a preset confidence threshold.

[0072] This set contains all detection candidates in the k-th frame that have passed the threshold test. According to a preset selection rule, a detection candidate is selected only if its confidence score meets the threshold requirement. Only then is the candidate deemed valid and retained.

[0073] III: When multiple high-confidence detection boxes point to the same license plate area in the image, the system needs to select the most representative single result to avoid duplication and waste of resources in subsequent processing.

[0074] The specific implementation is based on a non-maximum suppression algorithm. First, for the detection candidate set filtered by a threshold within a single frame, the system sorts the detection candidates in descending order of their confidence scores, ensuring that the processing order starts with the most reliable detection. The core of the algorithm lies in calculating the spatial overlap between any two candidate boxes, which is quantified by the intersection-union ratio (IU). In the formula, Represents bounding box and The intersection and union ratio of the fractions; Represents two bounding boxes to be compared; This is a function for calculating the area of ​​a rectangle.

[0075] The algorithm sets a threshold. The following iterative process is executed: select the candidate box with the highest confidence from the current candidate set and move it into the final output set. Subsequently, remove all candidates from the remaining candidate list that match the selected box for more than [a certain amount of time]. The candidate boxes are considered to highly overlap with the selected box and point to the same target. This process is repeated until the original candidate set is empty. The algorithm can be formally represented as: In the formula, It is the output of the NMS function, representing the final set of detection results after processing by the NMS algorithm; This represents the nonmaximum suppression algorithm function; This is the input to the NMS function, representing the result after filtering by the confidence threshold. The set of detection candidates retained in the frame image.

[0076] IV: For each frame in the deduplication result set The system iterates through each valid detection candidate. Based on the bounding box parameters, it selects the corresponding original image... The image cropping operation is performed. This step can be represented by the following formula as slicing the image matrix: In the formula, This indicates the first generation after truncation. Candidate images of license plate area; These are the ordinate and abscissa of the center point of the detection box in the original image coordinate system, respectively. These are the height and width of the detection frame, respectively.

[0077] After traversing and capturing all frames, the system will obtain a series of license plate area images. These images are associated with metadata such as detection confidence and timestamps from their sources. Finally, the system aggregates all the extracted images and outputs a high-quality candidate set of license plate images. It can be formally represented as:

[0078] Step S3 of this invention is license plate image quality assessment and selection. Specifically, the core task of this step is to systematically analyze and screen the candidate set of license plate region images from S2. By constructing a multi-dimensional quantitative evaluation system, the most suitable high-quality license plate image or image set for character recognition is identified and selected from multiple candidate images, thereby improving the accuracy and reliability of subsequent recognition processes.

[0079] Ⅰ: First, the candidate set Each image in Establish a multi-dimensional quality assessment system. Let the set of assessment functions be... Each function corresponds to a quality dimension: ① An improved Laplacian variance method is used to quantify the sharpness of image details. This method evaluates sharpness by calculating the response variance of the image after processing with the Laplacian operator: In the formula, For image Clarity score; Representing an image In pixel coordinates The pixel value at that location or the pixel value for each channel; The Laplace operator is a second-order differential operator; Indicates the image After applying the Laplace operator, at position The obtained response value.

[0080] ② Calculate the standard deviation of image pixel intensity to measure the distinguishability of license plate characters from the background: In the formula, Representing an image Contrast score; Representing an image The standard deviation of all pixel values ​​in the dataset.

[0081] ③ The integrity of the license plate character region is evaluated through edge detection and morphological operations. This metric calculates the proportion of the effective character region area to the total image area: In the formula, Representing an image Integrity score; This indicates the use of image processing techniques to extract information from images. The valid character region extracted from it.

[0082] ④ Use Hough transform to detect the angle of the main edges of the license plate and evaluate the horizontal alignment of the image: In the formula, image The tilt score; This indicates that line detection algorithms, such as Hough transform, are used to detect lines in an image. The angle between the main edge line identified in the image and the horizontal direction.

[0083] II: Due to the different units and ranges of the scores in each dimension, standardization is required to eliminate the influence of units. A min-max standardization method is used, processing each quality dimension separately. For the sharpness dimension, the maximum value of all images in the candidate set for that dimension is first calculated. and minimum value :

[0084] image The standardized resolution score is:

[0085] The system presets a set of weighting coefficients These correspond to four dimensions: sharpness, contrast, integrity, and tilt. (Image) The overall quality score is calculated by weighted summation:

[0086] III: Based on the calculated overall quality score, three configurable optimization strategies are provided, and the selected strategy is executed according to the preset mode parameters: ① Optimal selection strategy for a single frame: This strategy aims to select the frame with the highest overall quality as the representative image. The system iterates through the overall quality scores of all candidate images. Find the maximum value and its corresponding image index: This strategy directly outputs a single optimal image. It is suitable for scenarios that require high processing speed and assume the existence of a single frame of image.

[0087] ② High-quality collection selection strategy: This strategy sets a comprehensive quality threshold. Select all images that meet this quality standard to form a high-quality subset. The selection criteria are as follows: This strategy retains multiple high-quality candidates, providing possibilities for subsequent fusion of recognition results or multi-view verification, and is suitable for scenarios with extremely high requirements for recognition accuracy.

[0088] ③ Temporal Consistency Selection Strategy: This strategy, building upon the high-quality set selection strategy, further introduces a temporal continuity constraint. It not only requires that the image quality meets the standards but also that these high-quality images form a continuous sequence in the temporal dimension, ensuring they originate from the same stable phase of the vehicle's movement. The selection criteria include a quality threshold. and maximum time interval Two parameters: In the formula, For image The timestamps, and the images in the collection must be timestamped. After sorting, this condition is met. This strategy can effectively filter out isolated frames that, although of high quality, belong to different capture times and may lead to inconsistent recognition. It is suitable for tracking scenarios that require stable and consistent recognition results.

[0089] IV: Based on system configuration and application requirements, select one of the above strategies to determine the final output: As the final output of step S3, license plate image quality evaluation and selection, it is encapsulated and passed to the subsequent step S4. Regardless of Whether it contains one or multiple images, it represents the optimal input obtained after rigorous quality assessment and strategy selection, laying a reliable data foundation for subsequent character recognition.

[0090] Step S4 of this invention involves character segmentation and character-level quality perception. Specifically, the core task of this step is to refine the preferred license plate image from S3, segmenting it into independent individual character images, and performing independent quality assessment and classification on each character image, thereby providing structured input data and key quality prior information for subsequent differentiated character recognition.

[0091] Ⅰ: Set of license plate images The image is processed to enhance the features of the license plate area and to initially locate the regions where each character is located, providing a foundation for subsequent accurate character segmentation.

[0092] ① First, take each color license plate image Convert to grayscale image To reduce data dimensionality and computational complexity, a weighted average method is used for the transformation, taking into account the human eye's sensitivity to different colors. In the formula, Indicates coordinates The output grayscale image pixel values; These represent the input color images. In coordinates The pixel values ​​of the red (R), green (G), and blue (B) color channels at that location.

[0093] ② To highlight the difference between the characters and the background, the grayscale image was processed. Adaptive binarization is performed. This method can adapt to brightness variations in different regions of the image. For each pixel in the image... Within a local window centered on it, the mean and standard deviation of gray levels are calculated, and the threshold is dynamically determined. In the formula, Indicates coordinates The adaptive binarization threshold calculated at the specified location; Represented in pixels The average grayscale value of all pixels within a local neighborhood window centered on the pixel; The standard deviation of pixel grayscale values ​​within the same local neighborhood window; This is an adjustable parameter used to control the contribution of the standard deviation to the final threshold calculation.

[0094] Then binarization is performed: In the formula, To output a binary image in coordinates Pixel value at that location, Representing the future, i.e., the characters or the target. Represents the background.

[0095] For each pixel in the image, use its grayscale value. With an adaptive threshold specifically calculated for this point A comparison is performed. If the grayscale value is not lower than the threshold, the point is considered to belong to a brighter character area and is set as the foreground. Otherwise, it is considered to belong to a darker background area and is set as the background. Ultimately, the entire image was converted into a single image composed only of... and The resulting binary matrix clearly highlights the character regions as white blocks.

[0096] ③ Utilizing the characteristic that characters are horizontally arranged on a license plate, the binary image is processed. Perform vertical projection analysis. Calculate the number of foreground pixels (i.e., pixels with a value of 1) in each column of pixels: In the formula, Indicates the image number The vertical projection value of the column; Image height; Represents the binary image in coordinates The value at that location.

[0097] The drawn vertical projection curve Peaks will form in the column where the character is located, and troughs will form between the characters. By detecting the start and end positions of the troughs, the approximate boundary range of each character in the horizontal direction can be preliminarily determined.

[0098] II: Binarized image based on each selected image In addition, the preliminary horizontal boundary information of the characters obtained through vertical projection analysis is used to precisely delineate and extract each character region, accurately segmenting the license plate image into a series of independent individual character images.

[0099] ① For vertical projection curves The system smooths and verifies the trough locations, employing a sliding window averaging method to eliminate pseudo-troughs caused by minute fluctuations. In the formula, For in position The smoothed vertical projection value at the location; This indicates the width of the sliding window.

[0100] ② Within each horizontal interval determined by adjacent segmentation points, for the binary image Perform connected component analysis. Let the first... The first image The horizontal intervals are Extract the connected region consisting of all foreground pixels within this interval: Each character forms a primary connected component. The system selects the connected component with the largest area within this interval as a candidate target character and calculates its minimum bounding rectangle. : In the formula, Indicates the first The first image A bounding box for each character; The minimum x-coordinate among all pixels of the character; The smallest y-coordinate among all pixels of the character; The largest x-coordinate among all pixels of the character; It is the largest y-coordinate among all pixels of the character.

[0101] The rectangle precisely outlines the pixel-level position of the character in the image.

[0102] ③ Based on the calculated coordinates of the minimum bounding rectangle, extract the corresponding image block from the original preferred grayscale image to ensure that the original grayscale information is preserved. The extraction operation is a matrix slicing: In the formula, Indicates the extracted first... The first image The original image block of 1 character; For the first The grayscale image matrix of the license plate image.

[0103] Then, each extracted character image Normalization to a uniform size eliminates size differences in characters due to varying positions within the license plate or perspective during photography. Normalization employs bilinear interpolation. In the formula, This represents the number of normalized and scaled values. Zhang Image No. A standard-sized image of one character; This is an image scaling function; All character images will be scaled to this uniform size for the target height and width.

[0104] III: Perform multi-dimensional quality quantification analysis on each individual character image to provide accurate quality basis for subsequent character classification and differentiation processing. The evaluation focuses on sharpness, completeness, and positional rationality, which affect recognition accuracy.

[0105] ① The sharpness of characters directly affects the effectiveness of feature extraction. Therefore, a sharpness evaluation method based on gradient magnitude is first used to calculate the sharpness of character images. The gradient magnitude map, specifically used The operator calculates the gradient in both the horizontal and vertical directions: In the formula, These are the gradient component images of the character image in the x and y directions, respectively. They are respectively horizontal and vertical. Operator convolution kernels are used for edge detection.

[0106] Then, the gradient magnitude of each pixel is calculated: In the formula, For at pixel The gradient magnitude at a point represents the total intensity of the image intensity change at that point.

[0107] Character clarity score Defined as the average value of the gradient magnitude map: In the formula, This represents the total number of pixels in the character image. The higher the value, the sharper the character edges and the clearer the image.

[0108] ② Integrity assessment: Determines the integrity of the character structure, identifying any broken strokes or severe defects. This is achieved using a character template matching method. The system pre-stores a standard character template set, covering all possible characters. For the character to be evaluated... Calculate its normalized cross-correlation coefficient with all templates: In the formula, For character images and the first Template Normalized cross-correlation coefficients; This represents the average pixel value of the character image and the template image.

[0109] Character integrity score Take the maximum correlation coefficient: In the formula, Score the integrity of the characters; For all templates The correlation coefficient is taken at its maximum value; This is a collection of standard character templates.

[0110] ③ This assessment ensures that the relative positions of characters on the license plate conform to the specifications, preventing character misalignment due to segmentation errors. Based on prior knowledge of the license plate system, the position of each character... They all have an expected horizontal center position. and allowable deviation range Let the character The horizontal center coordinates of the original image are: Its location rationality score Calculated using the Gaussian function: In the formula, It is an exponential function.

[0111] This value reaches its maximum value of 1 when the character position is exactly as expected, and decreases as the deviation increases. A value close to 0 indicates that the position is extremely unreasonable.

[0112] IV: Based on the preset quality threshold, each character is classified into different states and labeled with a corresponding quality tag, thereby providing clear guidance for subsequent differential character recognition processing.

[0113] Specifically, the system is based on a sharpness score. and completeness score Two core metrics, combined with a preset sharpness threshold and integrity threshold Implement a three-way classification decision. Location rationality score. As an auxiliary verification indicator, when this value is abnormally low, the system will add a warning mark to the corresponding character, indicating that there may be a location error.

[0114] The logic of classification is fully defined by the following formula: The formula's judgment logic is as follows: First, check the integrity of the characters; if the integrity does not meet the standard... Regardless of clarity, it is directly classified as a missing character, because structural incompleteness is the most serious problem in recognition. If the completeness meets the standard, then the clarity is further checked: if the clarity meets the standard... The characters are classified as qualified; however, the clarity is substandard. Those are classified as fuzzy characters.

[0115] After completing this step, the system will output a clear quality label for each character. All characters and their labels together constitute the final output of step S4: a structured set of character images. and its corresponding quality label mapping This output fully describes the quality status of the license plate at the character granularity, providing an accurate quality judgment basis for step S5 to perform differentiated recognition based on character quality.

[0116] like Figure 3 As shown, step S5 of this invention is a differentiated recognition and feedback loop based on character quality. Specifically, the core task of this step is to allocate different recognition strategies and resources to characters of different qualities based on the character-level quality labels provided in S4, perform accurate recognition, and integrate the results; at the same time, a confidence-driven feedback mechanism is established so that when the confidence of the recognition result is insufficient, the preceding steps can be intelligently triggered to reprocess, thereby achieving robust recognition of difficult license plate characters such as blurred or missing characters at the system level.

[0117] Ⅰ: Receive the character image set from S4 and its corresponding quality label mapping Based on the quality label of each character, it accurately routes it to a preset recognition channel whose processing capacity matches its quality status, thereby achieving optimized allocation of computing resources and precise adaptation of recognition strategies.

[0118] The system first maintains three independent recognition processing channels, each consisting of a specific preprocessing module and recognition model. For a character image set... Each character image in The system reads its tag And perform deterministic routing decisions based on the labels: ① Standard Recognition Channel: This channel is dedicated to processing images labeled as valid characters. These images are clear and complete, requiring no additional enhancement. This channel is configured with a lightweight convolutional neural network recognition model. Its characteristics include fast reasoning speed and low resource consumption. Characters are directly input into the model for recognition: In the formula, To output tuples, It is the identified character category. This is the confidence score for the identification; This is a lightweight CNN recognition model function.

[0119] Meanwhile, the set of characters processed by this channel is denoted as: In the formula, The set of characters processed by the standard recognition channel.

[0120] ② Enhanced Recognition Channel: This channel is specifically designed for processing images labeled as blurry characters. These images have a generally intact structure but lack sufficient sharpness. This channel includes a character super-resolution enhancement function. and high-precision recognition model The processing flow is: enhancement followed by recognition. In the formula, For super-resolution enhancement function Processed character image; This is a high-precision recognition model function used to process the enhanced image.

[0121] Character super-resolution enhancement function Responsible for restoring the detailed textures of characters, high-precision model With a deeper structure and more powerful capabilities, it is responsible for making accurate judgments from the restored image. The set of characters processed by this channel is: In the formula, To enhance the character set processed by the recognition channel.

[0122] ③ Contextual Reasoning Channel: This channel is specifically designed for processing images labeled with missing characters. These images have severely incomplete structures, making reliable identification difficult based solely on image information. This channel employs a rule-based contextual reasoning engine. Its decision-making depends not only on the input image It also makes comprehensive use of the license plate coding rule knowledge base. Such as a list of province abbreviations, rules for alphanumeric combinations, and possible historical identification records. : In the formula, This is the output tuple for the context reasoning channel.

[0123] Inference Engine It may output a most probable character and its confidence score, or it may output a candidate list sorted by probability. The set of characters processed by this channel is: In the formula, The set of characters processed for the context reasoning channel.

[0124] II: Receive standard channel result set Enhanced channel result set Contextual reasoning result set The system also examines the correspondence between these results and the original character sequence. It reassembles the recognition results from different channels, which may be out of order, into a complete string according to the original spatial order of the license plate characters. Based on the recognition confidence of each character, it calculates an overall confidence index that reflects the reliability of the entire license plate recognition.

[0125] The specific execution process is divided into two stages: sequential assembly and confidence fusion. ① The system indexes each character based on its spatial position in the original license plate image. If the numbers are 1 to 7 from left to right, the recognition results scattered across different result sets are reordered. The assembly process uses a mapping function. Complete. The function takes three result sets as input and their mappings to position indices as input, and outputs an ordered sequence of tuples: In the formula, This represents the sequence of recognition results sorted by position. It is an assembly function used for merging and sorting; For a single result from a standard channel, include: character, confidence score, and position index; For a single result from the enhanced channel; This is a single result from the context channel.

[0126] Then, all character results Concatenate the strings sequentially to form the original recognition string. : In the formula, This represents a string concatenation operation; This represents the total number of characters in the license plate.

[0127] ② To assess the reliability of the overall license plate recognition results, Individual character confidence The results are merged into a single scalar value. The system uses the geometric mean as the fusion method because it is more sensitive to low-confidence values ​​and can better reflect the weak links in the overall chain. In the formula, Indicates the overall confidence level; This is the chain multiplication operator, representing multiplication by multiplication. From 1 to of Perform product operations; For the first The confidence level of character recognition.

[0128] Geometric Mean The value range is still between [0, 1]. The closer its value is to 1, the higher the confidence level of all characters and the more reliable the overall result. If the confidence level of any character is extremely low, it will significantly lower the overall value.

[0129] III: By comparing the overall confidence level with preset multi-level thresholds, a deterministic state determination is made on the result of the current identification attempt, thereby driving subsequent output or feedback decisions.

[0130] Specifically, the system presets three ordered confidence thresholds, forming two decision intervals. These three thresholds and their logical relationship are as follows: Success Threshold Retry threshold Failure threshold And all satisfy 0≤ ≤ ≤ ≤10≤ ≤ ≤ ≤1. The determination process is based on the overall confidence level. Comparison with these thresholds yields a clear ternary decision result.

[0131] The decision logic is formally defined using a piecewise function as follows: when A successful decision is triggered when the overall confidence level is high enough, the identification result is considered reliable, acceptable, and output. The system will then proceed to the final output process of S5-Ⅳ.

[0132] when The decision to retry is triggered when the overall confidence level is neither high enough for direct success nor low enough for complete failure. The recognition result has some uncertainty, but quality is improved through image restoration algorithms. The system will then enter the S5-Ⅳ feedback loop.

[0133] when The system will fail if the initial attempt fails. This decision indicates that the overall confidence level is too low, the current identification attempt is unreliable, and retrying may be ineffective. The system will then enter the S5-Ⅳ failure handling process, triggering manual review.

[0134] IV: Receive the judgment results from S5-III, including decision labels. Original recognition string and overall confidence level Its core task is to execute the final operation corresponding to the decision label and complete the entire license plate recognition process: if the recognition is successful, the result is output; if the confidence level is insufficient, a feedback loop is triggered to optimize the processing; if the recognition fails, a backup mechanism is activated.

[0135] The specific execution process involves branching based on decision labels, and the flow is as follows: The system maintains a feedback counter to record the number of feedback loops triggered in the current recognition process, and presets a maximum number of allowed feedback loops.

[0136] ①When A "successful" result indicates that the current recognition result meets the quality standards. The system executes: Original recognition string Used directly as the final identification result Output and the overall confidence level As the final confidence level Output all at once. Then, the process terminates normally, and the output is complete. .

[0137] ②When During a retry, the system first checks if the current value of the feedback counter is less than a preset maximum value. If the maximum number of retries has not been exceeded, the system will initiate a feedback loop mechanism: first, the feedback counter is incremented, and then a structured feedback instruction data packet is generated.

[0138] The data packet contains explicit reprocessing instructions, a reference to the original preferred image set, a summary of the current character quality analysis, and enables its backup image restoration algorithm to enhance the relevant images.

[0139] This image restoration algorithm employs a motion blur restoration model based on a convolutional neural network. This model learns a mapping from a blurred image to a sharp image through end-to-end training. For the input low-quality license plate image... The enhancement process is represented as: In the formula, This indicates a clear license plate image after enhancement and restoration. This represents a motion blur recovery model based on a convolutional neural network, with network weight parameters as follows: .

[0140] The training objective of this network is to minimize the following loss function: In the formula, This represents the total loss function during model training. The corresponding real and clear license plate image is used as the supervision target during training. For content loss, pixel-level mean squared error or L1 loss is typically used to measure the difference in pixel values ​​between the enhanced and sharpened images; To perceive loss, the similarity between the enhanced and clear images in deep semantic features is usually measured based on the difference in feature maps extracted by the pre-trained network. To counteract the loss, discriminator loss from the generative adversarial network framework is typically used, with the aim of making the enhanced image difficult to distinguish from the real, sharp image in terms of distribution.

[0141] The instruction is then passed to step S3, where the multi-dimensional quality score of the enhanced image is recalculated and selected again according to its quality assessment and optimization process. If the feedback counter has reached its maximum value, the system will not start a new loop, but will automatically determine the current state as a failure and transfer to the corresponding exception handling path.

[0142] ③When When a task fails or becomes unavailable due to exceeding retry limits, the system enters an exception handling process. In this path, the system sets the final recognition result with a specific failure flag and activates a manual review flag. Triggering this flag initiates a series of backup measures, including storing all relevant image data, intermediate processing results, and confidence analysis reports for the current task in a queue awaiting manual review, and notifying the system operator for intervention. Depending on the specific system configuration, this stage may also attempt to invoke a separate, more computationally demanding backup recognition service for a final attempt.

[0143] After completing this step, the entire intelligent license plate recognition process is finished. The system will then execute one of three paths based on the confidence level of the recognition result: directly outputting an accurate recognition result when the confidence level meets the success threshold; initiating an enhanced processing flow through a feedback loop when the confidence level is within an optimizable range; or triggering a manual review mechanism when the confidence level is too low. This confidence-based intelligent decision-making mechanism enables the system not only to output reliable results but also to possess self-optimization capabilities based on quality assessment, thereby significantly improving its robustness and practical value in complex scenarios.

[0144] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent license plate recognition, characterized in that, Includes the following steps: S1. Acquire multiple frames of images from the video stream, cache valid vehicle images after vehicle presence detection, and generate a multi-frame image sequence; S2, Use The network performs license plate region detection on a multi-frame image sequence. After confidence filtering and non-maximum suppression, the license plate region is extracted from each frame to form a candidate set of license plate region images. S3. Perform multi-dimensional quality assessment on each image in the candidate set of license plate area images, standardize the scores of each dimension, and calculate the comprehensive quality score by weighted fusion. Based on the score, select one or more high-quality license plate images from the candidate set. S4. Perform grayscale and adaptive binarization processing on the selected license plate images, locate character boundaries through vertical projection analysis, complete character segmentation to obtain individual character images, and conduct quality assessment on them, classifying characters into qualified characters, blurred characters, and missing characters. S5. Standard recognition is used for qualified characters, image-enhanced recognition is used for blurred characters, and contextual reasoning combined with encoding rules is used for missing characters. The recognition results of each character and their confidence scores are assembled into a complete license plate number in sequence, the overall confidence score is calculated, and a success threshold is set. If the overall confidence level reaches the preset success threshold, the complete license plate recognition result is output. If the success threshold is not reached, motion blur recovery enhancement based on convolutional neural network is performed on the original license plate image corresponding to the low confidence character, and steps S3 to S5 are repeated.

2. The intelligent license plate recognition method according to claim 1, characterized in that: The vehicle presence detection includes: inputting the acquired image into a trained binary classification network, which outputs unnormalized real values, and then... The function converts it into a confidence score for the presence of a vehicle and compares it with a preset decision threshold to determine whether the image contains a complete vehicle target.

3. The intelligent license plate recognition method according to claim 1, characterized in that: When caching valid vehicle images, a first-in-first-out (FIFO) queue structure is used to store and manage the valid vehicle images. When the queue is not full, new images are added in sequence. When the queue is full, the oldest image is overwritten and the read / write pointers are updated.

4. The intelligent license plate recognition method according to claim 1, characterized in that: The The network enhances its ability to detect license plate targets at multiple scales by introducing a feature pyramid network module into its structure, thereby optimizing the localization accuracy for small-sized and tilted license plates.

5. The intelligent license plate recognition method according to claim 1, characterized in that: Inputting the image Before processing the image, a fill scaling process is performed to maintain the aspect ratio, adjusting the image to the network input size and avoiding geometric distortion.

6. The intelligent license plate recognition method according to claim 1, characterized in that: When performing multi-dimensional quality assessment on each image in the candidate set of license plate area images, at least four dimensions of quantitative scores are included: sharpness, contrast, character region integrity, and image tilt. The sharpness is calculated using the improved Laplacian variance method, and the contrast is obtained by calculating the standard deviation of the image pixel intensity.

7. The intelligent license plate recognition method according to claim 1, characterized in that: When selecting high-quality license plate images based on the overall quality score, the overall quality scores are sorted in descending order, and one or more images with the highest scores are selected as the preferred results according to preset strategy parameters.

8. The intelligent license plate recognition method according to claim 1, characterized in that: The adaptive binarization process dynamically determines the binarization threshold of a pixel by calculating the mean and standard deviation of the grayscale value of a local window centered on each pixel.

9. The intelligent license plate recognition method according to claim 1, characterized in that: The classification of characters into qualified characters, ambiguous characters, or missing characters is a three-way classification decision made based on the clarity score and integrity score of each character, which are compared with preset clarity thresholds and integrity thresholds, respectively.

10. The intelligent license plate recognition method according to claim 1, characterized in that: The motion blur restoration and enhancement step includes: inputting a low-quality license plate image into a restoration model based on a convolutional neural network, wherein the model learns the mapping from a blurred image to a clear image through end-to-end training, and outputs an enhanced image for re-evaluation and optimization of quality.

Images